showcase: swap to moralstory fairness vector across all instruments
One mean_diff vector extracted from moral_stories_foundations fairness situations (not completions), calibrated once, read base/+c/-c on every instrument. On MFQ-2 it selectively raises the equality factor; on MFV it raises Fairness directionally but care leads. README reframed honestly, stale range_zoom/foundation_dlogit figures dropped (plotter no longer emits them). Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
@@ -17,9 +17,17 @@ with a human distribution over foundations (Clifford et al. 2015). Three configs
|
||||
`scifi`, `ai-actor`), two framings each (`other_violate`, `self_violate`).
|
||||
[[HF dataset](https://huggingface.co/datasets/wassname/tiny-mfv)]
|
||||
|
||||

|
||||
The bundled showcase steers Qwen3-4B with one activation vector, extracted once from
|
||||
`wassname/moral_stories_foundations` fairness situations (mean activation difference, fairness vs a
|
||||
balanced sample of the other foundations, situations only, no completions), iso-KL calibrated, then
|
||||
read across every instrument at base / +c / -c. On MFQ-2 it selectively raises the equality factor
|
||||
(the only factor that rises at the gentlest +c; it falls under -c); on the MFV vignettes it raises
|
||||
the Fairness foundation directionally but care leads, so the vector is fairness-leaning rather than a
|
||||
clean fairness isolator.
|
||||
|
||||

|
||||

|
||||
|
||||

|
||||
|
||||
## Install
|
||||
|
||||
@@ -159,7 +167,7 @@ Full return schemas live in the `TypedDict`s in `src/tinymfv/eval.py` and
|
||||
|
||||
## Steering plots
|
||||
|
||||

|
||||

|
||||
|
||||
`si_per_foundation` lives in steering-lite. It is a stricter, thresholded flip metric: did the steer
|
||||
move rows toward the intended foundation at `+C` and away at `-C`, while penalizing rows that were
|
||||
|
||||
|
Before Width: | Height: | Size: 188 KiB After Width: | Height: | Size: 528 KiB |
|
Before Width: | Height: | Size: 116 KiB After Width: | Height: | Size: 294 KiB |
|
Before Width: | Height: | Size: 271 KiB After Width: | Height: | Size: 266 KiB |
|
Before Width: | Height: | Size: 182 KiB After Width: | Height: | Size: 194 KiB |
|
Before Width: | Height: | Size: 392 KiB |
|
Before Width: | Height: | Size: 153 KiB After Width: | Height: | Size: 346 KiB |
|
Before Width: | Height: | Size: 89 KiB After Width: | Height: | Size: 194 KiB |
|
Before Width: | Height: | Size: 121 KiB After Width: | Height: | Size: 112 KiB |
|
Before Width: | Height: | Size: 78 KiB After Width: | Height: | Size: 79 KiB |
|
Before Width: | Height: | Size: 128 KiB |
|
Before Width: | Height: | Size: 187 KiB After Width: | Height: | Size: 391 KiB |
|
Before Width: | Height: | Size: 110 KiB After Width: | Height: | Size: 221 KiB |
|
Before Width: | Height: | Size: 112 KiB After Width: | Height: | Size: 103 KiB |
|
Before Width: | Height: | Size: 74 KiB After Width: | Height: | Size: 74 KiB |
|
Before Width: | Height: | Size: 132 KiB |
|
Before Width: | Height: | Size: 402 KiB After Width: | Height: | Size: 324 KiB |
|
Before Width: | Height: | Size: 220 KiB After Width: | Height: | Size: 184 KiB |
|
Before Width: | Height: | Size: 124 KiB After Width: | Height: | Size: 116 KiB |
|
Before Width: | Height: | Size: 78 KiB After Width: | Height: | Size: 80 KiB |
|
Before Width: | Height: | Size: 162 KiB |
|
Before Width: | Height: | Size: 65 KiB |
|
After Width: | Height: | Size: 112 KiB |
|
After Width: | Height: | Size: 66 KiB |
|
After Width: | Height: | Size: 95 KiB |
|
After Width: | Height: | Size: 60 KiB |